Amazon Bedrock Managed Knowledge Bases can now search CSV and Excel files by running calculations across their full tables, then combine the results with facts from related documents. That gives teams a way to ask multi-step questions across spreadsheets and policies without relying on a handful of retrieved rows to do the sums.
AWS AI Watch analysis
What happened
AWS says tabular search works through agentic retrieval, using the existing AgenticRetrieveStream request. When a question needs a table lookup or calculation, the system retrieves the relevant file, runs model-written code over the original in an isolated sandbox with no network access, and uses the results alongside document passages to form an answer with citations. The trace identifies the workbook the code read.
In AWS’s example, an assistant identifies departments that must write variance commentary by combining a finance policy, a quarter-close memo and calculations across a profit-and-loss workbook. Read AWS’s tabular search guide.
Why it matters
Ordinary retrieval can return a few promising spreadsheet passages while missing rows needed for a total, comparison or threshold check. Tabular search instead calculates across the retrieved file, which could make questions spanning business documents and workbook data more useful to answer in one place.
There is a trade-off: running code adds at least one agent turn, so answers using table data take longer than text-only answers. AWS says governed metrics, recurring dashboards and warehouse-scale joins still belong in business intelligence tools.
Our read
This is a meaningful step beyond asking a search system to find the right paragraph. It can work through the whole retrieved table, do the arithmetic and bring in the documents explaining what the numbers mean. That is useful for exploratory questions, not a reason to retire the finance team’s dashboards and controls.
What to watch
- Whether the system reliably retrieves the correct workbook, especially when draft and final files are similar.
- How users inspect the code-run trace and check calculations behind an answer.
- How response time and cost compare with dedicated spreadsheet or business intelligence tools.
Discussion spark: Would you trust an AI assistant to answer questions that combine spreadsheets and policy documents, or should calculations like these stay in dedicated analytics tools?
Sources and evidence
- Get answers from tables with tabular search in Amazon Bedrock Managed Knowledge Base (16 November 2026, 16:12 UTC)
- Use prompt caching with open-weight models on Amazon Bedrock (16 November 2026, 16:12 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)